Thales
← research journal

An internal research document, published verbatim by the automated daily export — not written for an audience, and better for it. All performance discussed is simulated paper trading; nothing here is investment advice.

COV-2 — the new coverage guard's threshold is keystone-pinned but its DENOMINATOR is a single-commit CSV that no guard watches, and the margin it is eating has no reader

status: open · raised panel run 49 (2026-11-09), change-reviewer lens, narrowed by adversarial review · class: registration hygiene (NORTHSTAR gauge G5) + loudness · judgement: YES · effort ~0.1 pd (leg 1) + ~0.1 pd (leg 3) · horizon: leg 1 bites immediately; leg 2 is a bounded measurement, not a deadline


Plain-language summary for an owner reading one paragraph

Six days ago the shop shipped a good guard: if too much of the price panel is missing its newest bar, refuse to select rather than trade on a partial universe (COV-1, #318 — the fix for the 2026-11-02 incident). Its threshold, 90%, is keystone-pinned, meaning a guard shouts if anyone changes it. But the number it is 90% of comes from data/universe.csv, a 903-name file that was committed once and that nothing pins and nothing refreshes. Editing or deleting that file silently moves a pinned threshold, and thales verify-config cannot see it. Separately, the guard writes its coverage margin into the order log and nothing displays it anywhere, so the one number that says how close the flagship is to refusing to trade is not on any report. It is currently at 98.0%, against a 90% floor — comfortable, but nobody is watching it and it has only moved one way.

Leg 1 — the pin and the thing it depends on are in different worlds (this is the reason to file)

  • execution.min_panel_coverage: 0.90 is in config_guard.KEYSTONES and MEANREV_KEYSTONES (added by #318). Good.
  • The denominator is expected = set(recent["symbol"].unique().to_list()) | set(universe or ()) (src/thales/execution/daily.py:116), where universe is _universe_symbols() (:905) reading config["universe"]["file"] = data/universe.csv.
  • config_guard pins universe.source: russell1000 (:62) — and that key is inert whenever the file exists, because load_universe (src/thales/data/universe.py:91) returns the file and consults source only when the file is absent. So the guard pins the key that does nothing and does not pin the key that decides the denominator.
  • data/universe.csv has exactly one commit — fa4ac852, 2026-08-18, carried incidentally inside a skew: snapshot 2026-08-18 commit — 903 symbols plus header, never modified since. (An earlier draft of this row cited dc5d457 / 2026-07-05; that was wrong, corrected here from git log -- data/universe.csv.)

So a one-line edit to an unwatched CSV moves a keystone-pinned safety threshold's effective strictness in either direction, with no guard complaint and nothing in verify-config --all-sleeves to catch it. That is the G5 shape — "% of consequential decisions whose criterion pre-dates the data" — with the criterion pinned and its basis unpinned.

Fix: pin universe.file, or pin the file's row count or content hash next to the threshold it serves; alternatively drop the inert universe.source pin in the same move so the pin set stops implying coverage it does not have. ~0.1 pd. Negative control: add a row to data/universe.csv and verify-config must fail; revert and it must pass. Today it passes either way (verified: verify-config --all-sleeves is clean with the file as committed, and nothing in it reads the file).

Leg 2 — what the margin is actually doing, stated as a limit on its own evidence

Measured from committed data only (data/raw is empty in a routine container, so this used data/pit_snapshots/*.parquet, which holds exactly the symbols that had a bar on each session). Universe symbols absent from a day's snapshot are that day's shortfall. "Persistently missing" = absent on every file in the window, with the four recorded truncated files (2026-08-27, 08-28, 10-12, 10-13) excluded so a 1-row file cannot dominate:

window (20 files each)persistently missing
2026-06-05 → 2026-07-072
2026-07-08 → 2026-08-077
2026-08-10 → 2026-09-1012
2026-10-06 → 2026-11-0515

On the newest session 18 names are missing, 15 of them from the persistent set: ASGN, AVB, BK, BLD, CTRA, EA, EQR, GTLS, HOLX, JHG, MASI, NSA, PSTG, TMHC, WBS — absent on all 40 most recent sessions, about ten weeks. Coverage today is 885/903 = 0.9801.

Three limits on this evidence, stated because they matter more than the trend:

  1. "Persistently missing over a 20-session window" is not "permanently dead." Several of the 15 are plainly live large caps — AvalonBay, Equity Residential, Bank of New York, Electronic Arts, Coterra, Hologic. So the mechanism is almost certainly "names the fetcher stops returning stay in the denominator", not index turnover, and a name could leave the set if fetching resumes. The trend is therefore not provably one-way.
  2. The terminal failure is far and this row does not rest on it. The refusal fires at 91 missing names (813/903 = 0.9003 passes, 812/903 = 0.8992 refuses); at the observed ~+2.6 names/month, with decelerating increments (+5, +5, +3), that is years away. This row does not ask anyone to act on that date — a fail-closed guard whose worst outcome is a loud refusal and a retry is not an emergency, and treating a 2029 projection as urgency would be the speculative-harm shape this shop kills.
  3. The denominator is deliberate and must not be changed. COV-1's body says "A panel of only 55 names scored 100%, which is why the universe is now in the denominator", and tests/test_execution/test_panel_coverage.py::test_cold_start_panel_cannot_hide_missing_names pins it. Nothing here proposes touching it.

What leg 2 is actually for: the live fact that the flagship has been selecting from 885 of 903 names for ten weeks and no detector anywhere says so. open/rwg-1 already names an AVB failed pair, so at least one of the 15 is a known-problem symbol. Whether "an R1000 name has had no bar in ten weeks" is worth its own detector is a separate question this row does not answer.

Leg 3 — the guard's own report-only fields have no reader

#318 added coverage fields written to data/processed/order_log.jsonl. Nothing renders them: grep for panel_coverage / n_ranked / n_held_unranked across src/thales/execution/{digest,fleet_digest,public_export}.py returns no output. So the margin exists as data and nowhere as information. One digest line would turn a cliff into a gradient, which is the cheap half of leg 2's point without any detector work.

Prior art — checked before filing

  • grep -ril over all four research/queue/ directories for universe.csv, denominator, min_panel_coverage, keystone, universe.source, universe.file; plus TECH_DEBT.md, AUDIT.md, RESEARCH.md, research/RESEARCH_QUEUE.md, web/public/data/killlist.json. No cov-2 slug anywhere.
  • built/cov-1-selection-runs-on-a-partial-panel.md is the parent: it states the design and the 98.1% / 4.1% threshold evidence, and never says universe.csv is unrefreshed or unpinned. #321's COV-1 annotation is entirely about DRO-1. So this is a residual of that build, not a relitigation of it.
  • The three queue files mentioning universe.csv (dismissed/skw-2, open/btb-1, built/stamp-1) use it as a symbol list for capture coverage — none is about it being frozen or being a guard's denominator.
  • Not on the kill list; proposes no trial, so no registry row and no deflated-Sharpe cost.

Test design

Leg 1: the negative control above (edit the CSV → verify-config must fail). Leg 3: assert the digest renders the coverage line, and that removing the renderer reds that test.

Pre-registered kill criterion

Drop this row, recording why, if any of:

  1. data/universe.csv acquires an automatic refresh, or the owner rules it a deliberately frozen artifact pinned elsewhere — then leg 1 is void.
  2. A measurement on the real store (needs data/raw, unavailable here) shows the 18-name shortfall is dominated by transient vendor misses rather than a stable set — then leg 2's framing is wrong and should be struck, though leg 1 survives independently of it.
  3. Leg 1 alone is judged below the bar. It is ~0.1 pd of pinning on a guard six days old; if the owner or steward reads that as tidying rather than registration hygiene, this row should die rather than sit open, and the measurement in leg 2 should be moved to a dated annotation on built/cov-1-*.

This row does not reopen on the margin shrinking further, and it does not reopen on the 91-name projection being revised: neither is its claim.

Where this ranks

Below the three live P0 rows (dro-1, rwg-1, gtr-1) and below open/pit-2, which has the only computed expiry in the tree. Leg 1 is the cheapest unpinning fix available and bites today; legs 2 and 3 are a measurement and a one-line report. Filing takes open/ to 13 against triage's post-sweep cap of 12, so this row is a candidate for overflow — and if triage trims it, the leg worth keeping in the dismissal is leg 1, not the measurement.

src/thales/utils/config_guard.py is not on ops/STEWARD.md §3's trading-path list, so leg 1 is observability-tier and needs no implementer slot.